product-readme
Authors public product README.md files: named host-run products (studio, harness, lab, domain kernel, agent system) get a category landing (mascot, category line, turn-host-into-job, counts, prompt); libraries get a short promise plus proof command. Then named techniques with limits. Use for OSS, CLIs, engines, and agent products. Do not use for internal platform-layer overviews (readable-readme) or package maps (extensive-readme). Unspecified "make a README" goes to the readme skill.
- License 许可
- MIT
- 最近更新
- 2026-09-15
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- product-readme
- description
- Authors public product README.md files: named host-run products (studio, harness, lab, domain kernel, agent system) get a category landing (mascot, category line, turn-host-into-job, counts, prompt); libraries get a short promise plus proof command. Then named techniques with limits. Use for OSS, CLIs, engines, and agent products. Do not use for internal platform-layer overviews (readable-readme) or package maps (extensive-readme). Unspecified "make a README" goes to the readme skill.
02 资源文件 3
- 参考文档checklist.md2.5 KB
- 参考文档examples.md2.4 KB
- 参考文档templates.md3.7 KB
03 SKILL.md 正文
Product README Authoring
Write a product landing a stranger can steal in two seconds. For a named product people clone and run with an AI host, that landing is a category claim, a host conversion, and counts from this tree. For a library, it is still what it is, what it is not, and a proof command.
Write for any installable product. Steal structure from public agent-studio landings (centered mascot, category one-liner, nav, conversion paragraph, prompt-first quick start). Do not copy their product names, slogans, or invented inventory. Do not invent a fourth README type.
Portable: never name a customer, private repo, or local disk path inside this skill. Describe the job. Fill every blank from the repo in front of you.
When to apply
- Installable product, OSS library, CLI, engine, or agent harness
- User asks for a product README, OSS landing page, or installable-library README
- Router default for installable / OSS / agent
Not this skill: internal platform layer / monorepo service → readable-readme.
Package maps → extensive-readme. Unknown type → readme (asks first).
If an extensive companion was requested, write README.md here, then load
extensive-readme. Banner (after nav, before the conversion paragraph):
> Full internals (every package, file map, how the repo runs): [Extensive README](docs/EXTENSIVE.md)
Workflow
Phase 1 — Discover
Do not invent features, numbers, or URLs.
- Kind: named host-run product (studio / harness / lab / kernel / agent system) vs library / SDK
- Narrow category a rival cannot paste; or skip "the first" if the category is crowded
- Host conversion: the tool they already run → the complete job
- Inventory counts you can point at in the tree (stop at two if a third would be padding)
- The prompt a stranger pastes, and/or the one command that proves it
- Primary interface, one invariant, one nearby-wrong
- 1–5 named techniques, each with an honest limit
- Existing logo / mascot; real docs / license the tree cites
Phase 2 — Length
| Shape | Length |
|---|---|
| Small SDK | Short: promise + 1–3 techniques + install |
| Named host-run product | Category landing: first screen + demos/proof + prompt-first quick start + techniques |
| Default | First-screen contract, skip empty sections |
Phase 3 — Write
Read ../copywriting/SKILL.md and
../readme/anti-slop.md first. Follow
templates.md. No ## 1. Vision numbering. Copywriting writes
the sentences. This skill owns section order. First-screen is / is not is
one factual nearby-wrong, after the conversion. Empty upgrade is banned.
Category claims with tree-true counts are allowed.
Phase 4 — Validate
Run checklist.md, the copywriting scorecard, and the anti-slop Pass.
Then critique the first screen as a stranger's screenshot:
- Can they steal the category sentence?
- Is there a "turn {host} into {job}" line?
- Are counts tree-true (no padded third)?
- Is the next action a prompt or a proof command, not a lecture?
- Did internals (ticket ids, harness letters) leak above the fold?
If any fail, rewrite the README. If the failure is a missing general rule, patch this skill (still no customer names), then rewrite again.
First screen (required)
Named host-run product, before the fold:
- Centered character mark (see Logo)
- Centered product name (H1). Do not duplicate it inside the mascot.
- Centered category line (
The first {narrow category}.in words a stranger already has; coined class names stay out of this line) - Real badges / nav only
- Extensive banner if a companion exists
- Conversion paragraph: turn {host} into {job}; plain-language input; named stages
- Count line from the tree (body; GitHub About may add "world's first")
- Blockquote: nearby-wrong, primary interface, invariant
- Proof: recorded demos / screenshots with teaching captions, and/or a fenced real command that does the job (
--helpis not the product check)
Library / SDK, before the fold:
- Logo if one exists (flat mark, not a mascot-by-default)
- Pitch: what it does in 1–2 paragraphs
- Blockquote: is / is not, primary interface, invariant
- Proof command
Get started / quick start for host-run products sits after proof and the workspace, still before a long technique catalog. Libraries may keep install after teaching. The proof command is never replaced by "see Get started".
Logo
Search assets/, docs/, docs/media/, public/, static/, brand/.
If a real logo exists: reuse it unless the user asked for a new mark.
If none, or the user asked for a mascot:
| Kind | Mark |
|---|---|
| Named host-run product | Character mark: small helper / companion, original vector, 2–3 colors. Domain cue as an attribute (bolt, lens, tool), not a pasted trademark. No photorealism. Transparent background. |
| Library / SDK | Flat geometric or short wordmark, 1–2 colors, no gradients / 3D / emoji |
Write assets/{product}-logo.svg (and a raster twin if the mark is a soft character). Mascot width in the README ~160–220. Lockup: character in the center, wordmark under it as a separate image in a real typeface (assets/{product}-wordmark.svg or .png). Do not use a default GitHub H1 as the name, and do not bake a long name into the mascot. Do not invent favicons. Do not copy a trademarked mascot.
Teaching
Each named technique:
- Memorable name
- Mechanism in plain language
- Honest limit (when it wins / loses)
- 1 verified link or omit
Optional field guide: vocabulary the reader can reuse. Cap at ~5 ideas.
Honest claims: only numbers this tree can reproduce. Point at a ledger or omit. Never invent tok/s, “trusted by”, or public-bench scores as fitness.
Section order
Named host-run product
- Mascot / name / category line / badges / nav
- Extensive banner if companion exists
- Conversion + count line
- Positioning blockquote (nearby-wrong / interface / invariant)
- Proof (demos, screenshots, or command)
- Try-these-prompts (pasteable, from this tree)
- Workspace / primary UI if any
- Quick start (prereqs, install, the prompt)
- How it works (short diagram) + named techniques with limits
- Optional field guide
- What it achieves (honest / omit)
- Go deeper
- License / contributing
Library / SDK: logo → pitch → blockquote → proof → why (if true) → techniques → get started → go deeper → license.
Anti-patterns
- Numbered Vision / Ideas skeleton on a product repo
- Generic Features with no idea
- Teaching essay before the category line on a host-run product
- Proof missing; install dumped before the reader knows what they installed
- Invented URLs, fake benchmarks, padded inventory counts
- File maps in
README.md - Internal codes on the first screen
- LLM default voice: ../readme/anti-slop.md
Additional resources
04 同仓库更多 Skills
greet
A greeting skill for testing
valid-skill
A valid skill for testing purposes.
unknown-field
A skill with an unknown frontmatter field.
no-frontmatter
暂无简介
child-skill
A nested skill in a subdirectory.
different-name
A skill with a name that doesn't match the directory.